1 s2.0 S0378873311000517 Main PDF

Social Networks 34 (2012) 82100
Contents lists available at SciVerse ScienceDirect
Social Networks
journal homepage: www.elsevier.com/locate/socnet
Geographical variability and network structure

Carter T. Butts a, , Ryan M. Acton b , John R. Hipp c , Nicholas N. Nagle d
a
Department of Sociology, Institute for Mathematical and Behavioral Sciences, University of California, Irvine, United States
Department of Sociology, University of Massachusetts, Amherst, United States
c
Department of Criminology, Law, and Society, University of California, Irvine, United States
d
Department of Geography, University of Tennessee, Knoxville, United States
b
a r t i c l e
i n f o
Keywords:
Spatially embedded networks
Spatial Bernoulli graphs
Geographical variability
Settlement patterns
Graph-level indices
a b s t r a c t
In this paper, we explore the potential implications of geographical variability for the structure of social
networks. Beginning with some basic simplifying assumptions, we derive a number of ways in which
local network structure should be expected to vary across a region whose population is unevenly distributed. To examine the manner in which these effects would be expected to manifest given realistic
population distributions, we then perform an exploratory simulation study that examines the features of
large-scale interpersonal networks generated using block-level data from the 2000 U.S. Census. Using a
stratied sample of micropolitan and metropolitan areas with populations ranging from approximately
1000 to 1,000,000 persons, we extrapolatively simulate network structure using spatial network models
calibrated to two fairly proximate social relations. From this sample of simulated networks, we examine
the effect of both within-location and between-location heterogeneity on a variety of structural properties. As we demonstrate, geographical variability produces large and distinctive features in the social
fabric that overlies it; at the same time, however, many aggregate network properties can be fairly wellpredicted from relatively simple spatial demographic variables. The impact of geographical variability is
thus predicted to depend substantially on the type of network property being assessed, and on the spatial
scale involved.
2011 Published by Elsevier B.V.
A basic observation regarding the distribution of humans across

geographical space is that this distribution is extremely heterogeneous. Even leaving aside the contrast between inhabited lands
and uninhabited oceans (comprising the majority of Earths surface area), settlements are typically concentrated in a small set
of regions having desirable geological, hydrological, and resource
access properties. Within these regions, the resulting settlements
are of extremely uneven size, distribution, and structure (Zipf,
1949; Brakman et al., 1999; White et al., 2008). Contrary to the
intuition of an evenly inhabited Earth, then, humans are distributed
unevenly across a wide range of geographical scales. Moreover,
this heterogeneity can be extreme. For example, the left panel of
Fig. 1 illustrates the population distribution of persons within the
Cheyenne, Nebraska metropolitan statistical area, as recorded by
the year 2000 Census of the United States.1 Although the population
This research was supported by NSF award BCS-0827027 and ONR award
N00014-08-1-1015.
Corresponding author at: University of California, Irvine, SSPA 2145, Irvine, CA
92697-5100, United States.
E-mail addresses: buttsc@uci.edu (C.T. Butts), racton@soc.umass.edu
(R.M. Acton), hippj@uci.edu (J.R. Hipp), nnagle@utk.edu (N.N. Nagle).
1
This and all maps shown are based on orthographic projections about a central point in the MSA, with distances in meters. Distance and area calculations
throughout the paper are based on these projections.
0378-8733/$ see front matter 2011 Published by Elsevier B.V.
doi:10.1016/j.socnet.2011.08.003
density in this area is approximately 3 persons per square kilometer (far below the U.S. mean of approximately 32 persons/km2 ), this
is not reective of the conditions under which most residents of
this region live. Indeed, as the gure indicates, most of the population of this region is concentrated into a small number of dense
communities, surrounded by large areas with few residents. The
extent of this concentration may be appreciated by comparing
the panel on the left with that on the right, which depicts a uniform population distribution over the same area. The difference
is stark. Rather than being embedded in a uniform, low-density
environment, the median resident of this region faces a local population density of approximately 1000 persons/km2 (as computed
from block-level data), with densities in some areas being as high
as 12,700 persons/km2 or as low as 0.09. The micro-environments
in which individuals form ties may thus differ greatly from a uniform baseline, and these micro-environments may themselves be
distributed unevenly across the landscape.
The above observations raise an important question: if human
settlement patterns are extremely heterogeneous, and if spatial
structure inuences network structure, then should we not expect
that the geographical variability in population distribution will
have a substantial impact on the structure of social networks? And,
if this is so, what will be the nature of that impact? To investigate these questions, we begin by employing a simple modeling
framework which preserves the marginal relationship between
C.T. Butts et al. / Social Networks 34 (2012) 82100
83
Fig. 1. Population distributions for Cheyenne, NE MSA. Points in left panel are placed uniformly within census blocks; points in right panel are placed uniformly over the
convex hull of positions from the census-constrained model. Both contain the same number of points (N = 9830).
distance and tie probability. Using this simplied framework, we

then derive a number of ways in which network structure should
(or, in some case, should not) be expected to vary based on the
underlying population distribution. To explore the ramications
of these results in a realistic geographical context, we perform an
extrapolative simulation study of network structure in communities from across the United States, simulating populations ranging
from 1000 to 1,000,000 persons in size over a range of land areas
using two relational models. By examining the properties of the
resulting networks, we draw some basic conclusions regarding the
likely impact of geographical variability on network structure, and
identify several concrete targets for empirical research.
Pr (Y = y|) =
ij B(Yij = yij |ij )), with parameter matrix given by
ij = F(Dij , ). Models of this form have been studied in the context of geographical distances by Butts (2002), Hipp and Perrin
(2009), Butts and Acton (2011), and are closely related to the latent
space models of Hoff et al. (2002), Handcock et al. (2007). They
can also be viewed as special cases of the family of gravity models (Haynes and Fotheringham, 1984), which have been used for
several decades in the geographical literature to model interaction between areal units. Butts (2006a) has further shown that the
spatial Bernoulli graphs can be written as a special case of a more
general curved exponential family of graph distributions. By dening canonical parameters (, d) = logitF(d, ), we may write the
pmf for adjacency matrix Y with support Y as
1. Spatial models of social networks

It is a well-established result that the marginal probability of
a social tie between two persons declines with geographical distance for a wide range of social relations (see, e.g., Bossard, 1932;
Zipf, 1949; Festinger et al., 1950; Hgerstrand, 1967; Freeman et al.,
1988; Latan et al., 1995; McPherson et al., 2001). While often
regarded as a mere curiosity, others have argued that this relationship is a critical determinant of social structure (Mayhew, 1984b).
Indeed, Butts (2003) has shown that under fairly weak conditions,
spatial structure is adequate to account for the vast majority of
network structure (in terms of total entropy) at large geographical
scales.
The simplest family of network models to incorporate this
notion is the family of spatial Bernoulli graphs, dened by probability mass functions (pmfs) of the form
Pr(Y = y|D) =
B(Yij = yij |F(Dij , )),
(1)
{i,j}
where Y is the (random) graph adjacency matrix, D is a matrix of

inter-vertex distances, B is the Bernoulli pmf, and F is a function taking distances into the [0, 1] interval (parameterized by real vector
). In this context, F is referred to as a spatial interaction function (SIF), and can be interpreted directly as providing the marginal
probability of a tie between two randomly selected individuals at
some given distance. It can immediately be observed that this family is a special case of the inhomogeneous Bernoulli graphs (w/pmf
Pr(Y = y|D, ,
) exp[
(, Dij )yij +
t(y)],
(2)
{i,j}
where Rp and t : Y Rp are respective vectors of parameters

and sufcient statistics. The incorporation of additional statistics
(via t) allows for the combination of both spatial and non-spatial
effects (e.g., endogenous triangulation, as explored in recent work
by Daraganova and Pattison (2007)).
From the perspective of this general family, we re-obtain the
spatial Bernoulli graphs in the case for which all non-spatial effects
are omitted. In practice, this is naturally an approximation (though
possibly a good one for large-scale structure); however, it also has
an important theoretical interpretation as the graph distribution
having maximum entropy given the marginal distance/tie probability relationship. Thus, if one seeks to understand the pure impact
of spatial structure on network structure, this model serves as a
natural reference point. This model also has the attractive property
that it is amenable to large-scale simulation (something not true
of models with more complex effects see, e.g. Snijders (2002)
for a discussion of some relevant difculties), and in some cases
to analytical treatment. Given its simple interpretation and theoretical leverage, we employ the spatial Bernoulli framework in the
work that follows; we do note, however, that the ability to extend
it via the method of Eq. (2) provides a ready avenue for further
exploration.
84
2. Effects of spatial heterogeneity: intuitions from rst

principles
To reiterate the motivating observation of Fig. 1, human beings
are heterogeneously distributed over the Earths surface. If, per
Section 1, the probability of a social tie between two persons varies
systematically with distance, then this heterogeneity in spatial
structure will be reected in the structure of the associated social
network. Although some effects of spatial heterogeneity2 on network structure are difcult to characterize, it is possible to obtain
an intuition for some of the relevant mechanisms by examining a
few stylized scenarios. In this section, we consider several such scenarios, with the goal of highlighting some ways in which variation
in population density would be expected to affect the structure of
social networks.
To begin, let us consider some xed region, A, to which we
sequentially add vertices; we assume that the location of each new
vertex is chosen by some iid process, and that ties from old to new
vertices are governed by the model of Eq. (1). For any xed vertex,
vi , the probability of a tie to a newly added vertex, vj is given by
Pr(Yij = 1|i , ) =
p(j )F(Dij , )dj pY (i , )

A
where i , j are the locations of the respective vertices and p( j )

is the population distribution over space. Note that, by the iid
assumption, the above works for any vertex, vj , placed after vi .
Moreover, since all vertices are placed independently, we can without loss of generality take vi = v1 to be the rst vertex placed.
Then, after k vertices are added to the graph, the conditional local
expected degree of vi i.e., the expected degree within the subgraph
induced by membership in A is simply
k
Y1j |1 =
j=2
k

j=2
EY1j |1 =
k
pY (1 , ) = kpY (1 , ).
j=2
Marginalizing over 1 gives the expected degree of vi ,

k
j=2

Y1j =
kpY (1 , )d1 ,

A
which again by the iid assumption reduces to k times a function that depends only on the xed vertex location pdf and the
the above right-hand expression can be rewritten as
SIF.
Indeed,

k A A p()p()F(D, , )dd, with the double area integral being
the marginal probability of an edge between two randomly selected
vertices.
This simple exercise leads to an observation that we may call
the in-lling principle: adding vertices to a xed region in a
uniform way leads to a linear increase in expected local (withinregion) degree, while holding the expected local density constant.
(This last follows immediately from the well-known graph identity
= d/(N 1), where is the density and d is the mean degree.) Likewise, given two regions of equivalent shape and area and the same
local population distribution, we expect the ratio of their internal
mean degrees to be equal to the ratio of their population sizes (with
their internal densities again being equal). Where the population
gradient is relatively uniform, we thus expect local mean degree to
scale linearly with population density.
2
Throughout this paper, we will use the term spatial heterogeneity to refer
generically to variation in local population density across space. Other forms of
heterogeneity are also possible (e.g., non-stationarity of the SIF), but we focus here
on this particular case.
The in-lling principle has a number of further consequences,

which may be appreciated by considering the impact of increasing mean degree on other structural properties. As average degree
increases, the probability of connectivity rises; for d ln N, almost
all graphs are connected (see, e.g., West, 1996, pp. 413417), and we
may thus expect that smallregions
will become locally connected

where ln N/N
, with = A A p()p()F(D, , )dd being the
local expected density. (This threshold is not exact for non-uniform
F due to spatial clustering, but does serve as a lower bound.) This
effect is illustrated in the left-hand panel of Fig. 2, which shows the
probability of a regional networks being connected as a function
of N and in the uniform case (red and green curves). Increasing N at constant network density also increases the probability
of higher order connectivity (e.g., biconnectivity), leading to local
populations that are robustly connected. The population threshold
for biconnectivity is somewhat higher than that for connectivity
(see Fig. 2, yellow and cyan curves), but its behavior is otherwise
similar.
The practical import of these observations is illustrated
schematically in Fig. 3. For an arbitrary region of constant shape
having area , we can identify some threshold population density
() such that uniform placement of () vertices within such a
region will lead to an expected mean degree exceeding the connectivity threshold for SIF F. Cutting the population density surface
at () will reveal regions expected to be locally connected; specifically, induced subgraphs taken by selecting all vertices belonging
to appropriately shaped regions of area in the regions above the
threshold are expected to be connected with high probability. Similar subgraphs taken from below-threshold regions are unlikely to
be connected. Since the union of two non-disjoint connected subgraphs must itself be connected, the above also implies that large
regions exceeding the threshold will tend to be globally connected (with connectivity extending to at least the edge of the
above-threshold region). Thus, the in-lling principle leads us to
expect local social connectivity to vary systematically with the population density surface, and, moreover, for there to be a fairly clear
spatial boundary between connected and disconnected areas.
In addition to connectivity per se, another property potentially
impacted by in-lling is local cohesion. Many notions of subgroup
cohesion exist (see, e.g., Wasserman and Faust, 1994, for a review),
with most considering subgroups to be cohesive to the extent that
they are locally dense, that members are not easily disconnected
from the group, and/or that group members are socially proximate
to one another. To take a simple example, the k-core of a graph
is dened to be the maximal set of vertices such that all vertices
within the set are adjacent to at least k other set members. Although
the k-cores are globally cohesive in only a minimal sense even
high-order cores need not be connected, for instance they can
be seen as unions of well-connected groups. In particular, every
member of the 1-core belongs to a connected component of size
greater than 1, every member of the 2-core belongs to a biconnected component, etc. The number of the highest-order k-core
to which a vertex belongs, then, is a convenient indicator of the
extent to which it is embedded in a well-connected group (even
if the core as a whole is poorly connected). We shall refer to this
number henceforth as the core number of a vertex. At constant
density, the formation of higher-order cores and hence higher
core numbers is strongly associated with graph order. The righthand panel of Fig. 2 shows the development of core structure in a
random graph of constant density as order increases. Initially, vertices transition from the 0-core (i.e., isolates) to the 1-core, with this
process slowing as the 2-core begins to emerge. (Core membership
as depicted here refers to the highest-order core, or core number.)
This process is repeated as N grows, with successive cores emerging and being consumed by higher-order cores at regular intervals.
Averaging across the population, however, we see that the mean
85
Fig. 3. Emergence of local connectivity on an uneven population density surface.

Where the threshold population density for an approximately uniform region of
area is () (such that () ln N/N), local connectivity emerges where dN/dx
exceeds the threshold for intervals of appropriate area. Similar intervals of lower
population remain locally disconnected.
Fig. 2. Effects of increasing order on connectivity and cohesion for random graphs
of xed expected density. Left panel shows probability of connectivity and biconnectivity by order and density. Right panel shows fraction belonging to each k-core
(and no higher) and mean core number (dotted line) at 1% expected density, by
graph order.
core number (black line) rises steadily, growing approximately linearly with N. As with connectivity, these behaviors may be altered
somewhat by spatial clustering. They provide a useful intuition,
however, for the baseline impact of in-lling on local structure.
Taking the strong relationship of mean core number with N at

constant network density together with the in-lling principle, we
would expect that regions of a given shape and area having higher
local population densities will exhibit higher mean core numbers
than equivalent regions of low population density (again, in the
sense of Fig. 3. Since the scaling of mean core number with N is
roughly linear (in contrast with the threshold behavior of connectivity), we also expect that variation in mean core number across
the population surface will be much smoother than variation in
local connectivity. At the same time, membership in cores of a particular order is likely to be rare until a particular population density
threshold is reached (with that threshold varying depending on the
order of the core in question). For phenomena expected to emerge
only in subgroups of a particular level of cohesion, then, we may
still expect qualitative shifts in behavior between high and low density regions. For phenomena assumed to vary quantitatively with
local cohesion, direct proportionality to the local population seems
a reasonable rst guess.
Although considerable insight can be gleaned from studying
stylized situations, the arguments of this section omit several
potentially consequential factors. As noted earlier, clustering due to
spatial effects should tend to increase the N needed for connectivity,
relative to a homogeneous Bernoulli model with the same expected
density. We can bound this effect as follows. Let G be the spatial
Bernoulli graph on the vertices of region A associated with SIF F.
Dene = min,A F(D, ), and let G be a homogeneous Bernoulli
graph with parameter on the same vertex set. Intuitively, every
pair in A is adjacent with probability at least , and G can be thought
of as supplying a lower bound on the true graph G.3 Now, dene
the residual Bernoulli graph R, with parameter matrix such
that ij = 1 (1 F(Dij ))/(1 ). Let G R represent the union
of G and R, i.e. the random graph in which an edge appears iff it
appears in G , R, or both. Since all edges of G R are independent
and occur with probability 1 (1 )(1 ij ) = F(Dij ), it follows
that G G R.
This decomposition of the spatial Bernoulli graph G into a
homogeneous Bernoulli graph G and a residual graph R can provide us with a rigorous tool for understanding the behavior of
G in more complex settings. For instance, if some draw g from
G is connected, then g R is connected; thus, the probability
of connectivity under G is a lower bound on the probability of
connectivity under G. More generally, let z be any graph statistic
such that z(x y) z(x) for graphs x and y having the same vertex set. Then clearly Ez(G) Ez(G ) and, for any given value zo of z,
Pr (z(G) zo ) Pr (z(G ) zo ). Since many statistics of interest (e.g.,
3
E.g., we can generate G and G using the same random inputs, such that every
edge in G is also in G ; see Butts (2010) for a general treatment of this approach.
86
mean core number, probability of k-connectivity) satisfy the condition of z, it follows that we can employ the properties of G to bound
the behavior of G. On the other hand, not all properties are preserved under graph union (e.g., betweenness centralization), and
the bound obtained in some cases may be too loose to be useful
(e.g., if
). Thus, this is an incomplete solution.
Another factor ignored thus far has been the role of non-local
ties. Clearly, it is possible to identify some region A whose induced
subgraph is disconnected with high probability, while the induced
subgraph of some larger region B A is very likely connected
consider, for instance, the case of vertices distributed at unit
intervals on the real line, with A being a segment of length 2, B
being a segment of arbitrarily long length , and F a constant
k such that ln /
k
1. Although this example is implausible,
it points to a real phenomenon: particularly when F is heavytailed, the locally induced subgraph for a small region may not
effectively characterize the properties of its members within the
larger network. This effect will itself vary across the population
surface, to the extent that high-population regions will tend to be
embedded within or adjacent to other high-population regions, and
low-population regions will likewise be associated with other lowpopulation regions. The exact impact of this effect will vary with
the model SIF, graph statistic examined, and population surface,
and is difcult to characterize on an a priori basis.
Considering simplied scenarios gives us general insights into
the rst-order effects of spatial heterogeneity on network structure,
but does not tell us how these factors will play out in practice. To
move beyond these basic intuitions, we must examine the behavior
of spatial Bernoulli graphs under realistic conditions: that is, with
SIFs based on real data, and vertex positions that are reective of
geographical reality. It is to this problem that we now turn.
3. Bringing geography back in: a simulation study
While the arguments of the previous section suggest various
general ways in which network structure should vary across space,
they also underscore the fact that such effects are contingent on the
underlying population surface: the same mechanisms can lead to
very different networks when applied to populations that are differently distributed. What would we expect to see in real networks,
then, given empirically observed settlement patterns? Simulation
provides a natural means of addressing this problem. Specically,
we may utilize detailed data on population distributions to simulate draws from spatial Bernoulli graphs with xed SIFs, and analyze
the resulting networks to examine the impact of spatial structure
on network properties (holding other factors constant); it is this
approach that we employ here. Although the models we employ
are empirically realizable (and our simulations based on extrapolations from observed data, rather than rst principles alone), it
should be borne in mind that our purpose is to explore the theoretical implications of spatial structure for network structure given
a minimal set of assumptions, as opposed to making detailed predictions about particular cases. In this, sense, this effort lies at a
midpoint on the intellective versus emulative modeling continuum discussed by Carley (2002), incorporating certain aspects of
empirical detail while retaining enough simplifying assumptions to
permit a relatively general analysis. At the same time, the generality
of the framework in which our models are embedded (i.e., the spatial ERGs of Eq. (2)) permits subsequent extension and elaboration
where appropriate.
3.1. Simulation design
Our simulation study proceeds as follows. First, we choose a set
of locations to examine, selecting them so as to evenly cover the
Fig. 4. Population size and land area distribution, U.S. micro and metropolitan statistical areas. Vertical and horizontal lines indicate sampling strata for this study;
selected locations are those closest to line intersections.
range of observed population sizes and land areas for U.S. micropolitan and metropolitan areas (as well as a number of other, smaller,
inhabited locations). For each location in this set, we then simulate
population microdistribution using data on block-level population
counts and household size distributions. Finally, we simulate networks using two previously calibrated spatial Bernoulli models for
each location. By analyzing the resulting set of networks, we are
then able to examine the impact of within and across-location geographical variability on network structure.
3.1.1. Location selection
To compare network structure across regions, one would ideally like to employ units that are both well-dened and socially
bounded. While few places in the developed world are fully isolated
from one another, it is nevertheless possible to identify regions that
correspond to relatively well-dened social units with respect to
such processes as daily migration, employment, local commerce,
and everyday interaction. Using such criteria, the U.S. Department
of the Census divides the populated regions of the United States into
micropolitan and metropolitan statistical areas, each of which contains one or more towns, cities, or other agglomerations together
with the immediately surrounding area (to the nearest county
or parish boundary, as applicable). A metropolitan area contains
at least one city with a population of at least 50,000, whereas
a micropolitan area contains at least one city with a population
between 10,000 and 50,000. An area in this sense includes the
primary city and the surrounding county; it also includes adjacent counties if they are highly socially integrated based on journey
to work patterns. At a minimum, micropolitan/metropolitan areas
are thus collections of locally dense population surrounded by a
low-population buffer, although the Census denition also seeks
to avoid splitting areas experiencing strong interaction in other
respects. The unied set of 922 micropolitan and metropolitan
areas, along with the other 1379 smaller counties in the United
States, serve as the population for our study, with the individual
area (henceforth location) being our primary unit of analysis.
In order to cover areas with a wide range of geographical properties, we employ a stratied sample of locations selected by
population size and land area (respectively). As location sizes on
both dimensions scale roughly logarithmically, we identied four
target strata on each dimension based on the overall distribution of
locations. (See Fig. 4.) These stratum values were 103 , 104 , 105 , and
106 for population size, and e7 km2 , e8 km2 , e9 km2 , and e10 km2
for land area. For each pair of population size and land area values, the location was identied whose population and area was as
close as possible to the target (in a least squares sense), yielding
87
Table 1
Sample locations, stratied by population and land area.
Population stratum
Location
1 10
Bristol Bay Borough, AK

Golden Valley, MT
Esmeralda, NV
Yakutat, AK
Choctaw, MS
Cheyenne, NE
Quay, NM
White Pine, NV
Lawrence, KS
Cookeville, TN
Idaho Falls, ID
Navajo, AZ
Honolulu, HI
Hartford, CT
Rochester, NY
Salt Lake City, UT
1 104
1 105
1 106
Population
1258
1042
971
808
9758
9830
10,155
9181
99,962
93,417
101,677
97,470
876,156
1,148,618
1,037,831
968,858
a total of 16 study locations. These are listed in Table 1. As can

be seen, the selected locations run the gamut from rural and even
near-wilderness locations to long-settled urban environments. By
examining the properties in predicted network structure across
this range of environments (holding constant other factors), we
may assess the extent to which geographic variability would be
expected to drive structural properties throughout the population
of U.S. micro and metropolitan areas.
3.1.2. Population microdistribution
For purposes of detailed network simulation, it is necessary to
associate each individual within each location with a particular
point in space. Such point placement can be immediately constrained by Census counts, which identify total population at the
block level (an areal unit roughly analogous to a city block in urban
areas, constructed on the basis of physical boundaries, population
homogeneity, and population size). Within blocks, most individuals can be further localized to households, whose size distribution
within each block is also known. This leaves the placement of
households and the placement of persons within households. In
this paper, we refer to the resulting spatial distribution of individuals within blocks as the population microdistribution. To simulate
the microdistribution, we utilize two competing models representing the extreme cases of a range of potential point processes. The
rst assumes a maximum entropy solution, in which households
(or isolated individuals) are placed uniformly at random subject to
known geographical constraints. The second, by contrast, assumes
a near-minimal entropy solution, in which households are placed
in an extremely even, grid-like manner using a low-discrepancy
sequence (specically, a two-dimensional Halton sequence4 ). We
refer to these microdistribution models henceforth as uniform
and quasi-random placement, respectively. By examining the
impact of these extreme cases on social structure, we can thus
infer the likely range of possibilities which could be occupied by
processes having intermediate behavior. For both models, we also
avoid unrealistic ground-level congestion by means of a simple articial elevation model, which simulates the effects of multi-story
residential structures in densely populated blocks. Specically,
households whose ground position would place them within a
10 m radius of k previously placed households are given a vertical elevation of 4k meters; thus, intuitively, articial elevation
4
A Halton sequence is a deterministic sequence of points that lls space in
a uniform manner, while also maintaining a high nearest-neighbor distance. The
result (sometimes called a quasi-random distribution) is similar to a set of draws
from the uniform distribution, but substantially more evenly placed; see Gentle
(1998) for algorithmic details.
Land area (km2 )

1308
3044
9294
19,815
1085
3099
7446
22,989
1183
2961
7676
25,779
1553
3923
7592
24,705
Population density (Pop/km2 )

0.96
0.34
0.10
0.04
8.99
3.17
1.36
0.40
84.48
31.55
13.25
3.78
564.03
292.77
136.69
39.22
arises as population density grows, with new households stacking

on old ones. (Arrival order is treated as random.) Finally, withinhousehold proximity is maintained by requiring household size to
satisfy the known marginals within each block, and then placing
individuals at their household locations (jittering randomly within
a 5 m radius to avoid exact overlap).
An example of the result of this microsimulation process is
shown in Fig. 5. Both panels depict the Quay, New Mexico MSA, with
lines designating the census block boundaries (2690 in all). Vertex
locations (N = 10, 155) are indicated by colored dots, with articial
elevation denoted by dot color (ranging from red, at ground level, to
violet at 20 m). As can be appreciated from the large areas of empty
space, most of the population is heavily concentrated into a relatively small fraction of blocks. Thus, population layout is to a great
extent constrained by the available data. The difference between
uniform (left panel) and quasi-random (right-panel) microdistribution models is seen in the relative evenness with which population
is distributed within each block: a process of uniform (random)
placement leads to numerous clumps in the household distribution, a phenomenon that is minimized under the quasi-random
process. An immediate side effect of the latter is that the quasirandom process (approximately) maximizes the distance between
households, a property that could be reasonably anticipated to
affect network cohesion. Another side effect is a reduction in articial elevation, stemming from the more efcient use of space under
the quasi-random process. This is readily apparent in Fig. 5 insets,
which show an expanded view of a 2 km square portion of the city
of Tucumcari, NM. In this relatively high-density region, both models predict that some degree of articial elevation will occur; due
to the higher rate of clumping in the uniform model, however,
the later produces a larger number of elevated households. While
our intent in this simulation is merely to produce a plausible level
of inhibition for ties among individuals in high-density location
(rather than to correctly capture the true distribution of residence
heights per se), we note that the elevations produced by this process on our 16 locations are broadly consistent with the ranges of
residential building heights typically reported in the literature (e.g.,
Burian et al., 2002).
3.1.3. Spatial interaction functions

For marginal models, the effect of geography on social interaction is captured through the spatial interaction function. While
many choices of SIF are possible, our interest in this case is in
functions that plausibly represent families of relatively proximate
relations. To that end, we employ two specic functions obtained
in prior work by Butts (2002) based on analyses of existing data
sources. The rst is from a social friendship relation (based on
88
Fig. 5. Comparison of uniform and quasi-random vertex placement, Quay County, NM MSA. Lines indicate census block boundaries, with articial elevation shown via vertex
color. Insets provide detail of 2 km 2 km portion of Tucumcari, NM. (For interpretation of the references to color in this gure legend, the reader is referred to the web
version of the article.)
data from (Festinger et al., 1950)), and can be thought of as a

locally sparse relation with a fairly long tail (declining as approximately d2.8 for large distances). The second is from a face-to-face
interaction relation based on data from Freeman et al. (1988),
which acts as locally dense relation that attenuates very quickly
with distance (apx d6.4 ). Both are of the general power law form
F(x) = 1 /(1 + 2 x)3 , with parameter vectors (0.533, 0.032, 2.788)
in the case of the social friendship model, and (0.859, 0.035, 6.437)
in the case of the model for face-to-face interaction. While we do
not assume that all networks or, indeed, that all types of friendship or interpersonal interaction follow one of these two forms,
we take these as plausible examples of the types of SIFs one is
likely to see from proximate relations such as those from which
the functions were derived. Similarities and differences in the networks formed by such functions thus provide us with some sense
of the range of behaviors one might observe in similar settings.
In computing distances for purposes of the SIF, we employ the
Euclidean distances between individual surface positions in the
projected geometry, plus a small correction to account for the
effects of the built environment (where applicable). Specically,
differences in articial elevation are added to the base Euclidean
distance for those within 25 m of one another, while the sums
of individuals articial elevation distances are added for pairs
whose surface positions differ by more than 25 m. This simulates a
basic feature of travel within the built environment, namely movement within a building for those who are otherwise proximate,
versus movement to down ground level, over to the second position, and up for those with distant surface locations. While one
could employ more complex schemes (including explicit adjustment for roadways, local obstructions, etc.), this would require
more detailed knowledge of household position, built environment, and other aspects of local geography than were available
for our test regions. As a practical matter, experimentation with
reasonable alternatives to the approach employed here did not
produce substantively different results. By way of explanation,
it should be noted that different notions of distance tend to be
very strongly correlated, even at fairly small scales. For instance,
a comparison of Euclidean distances for the cases used here with

Manhattan distances (the so-called city block metric sometimes
suggested as an alternative for urban environments) produced a
median correlation of approximately 0.99 under both uniform and
quasi-random microdistribution models; even considering only
very proximate points (within 100 m of each other in Euclidean
space), the median correlation is still approximately 0.98 for both
microdistribution models.5 While the impact of alternative distance models on network structure at smaller spatial scales is an
interesting and potentially productive target for further research,
our experiments suggest that the results reported here are not sensitive to reasonable variations in how distance is dened.
3.1.4. Network simulation
Given the above, simulation proceeds as follows.6 For every
location, we generate a population microdistribution using each
placement model (uniform and quasi-random), subsequently generating 25 networks from each microdistribution using the two
spatial models. The resulting set of 1600 networks is the primary
basis for our subsequent comparative analysis. In addition to this
larger sample, an additional single network was drawn in each
condition. This smaller set of 64 networks was retained for withinnetwork analysis.
3.2. Results
Our analysis of the simulated networks includes an examination of both within and between location heterogeneity. Here, we
5
Since Manhattan distance is affected by the choice of coordinates, we also
considered the correlation of Euclidean distance against Manhattan distance on a
randomly rotated axis set. The resulting median correlations were nearly identical
(apx 0.99 in the unconstrained case, and 0.97 within 100 m).
6
Simulation and analysis was performed using the statnet and sna libraries for
R (Handcock et al., 2008; Butts, 2008) and the R spatial tools (Bivand et al., 2008),
along with additional functions created by the authors.
discuss our major ndings in each area, beginning with crosslocational comparisons.
3.2.1. Graph-level properties across locations
Given our widely dispersed set of study locations, and the high
level of spatial heterogeneity within each location, it seems plausible that few systematic patterns will be found that span the
sample of simulated graphs. If, on the other hand, we nd that
there are aggregate properties that remain stable or that change
in a predictable way across locations, then we may provisionally conclude that these properties will be robust enough to justify
empirical investigation (for networks with similar SIFs to those
considered here, at least). In this section we consider several contexts in which graph-level relationships can occur: comparisons of
spatially conditioned and uniform random graphs; associations of
graph-level properties with aggregate geographical features; and
correlations among multiple global properties on the same networks.
3.2.1.1. Comparison with random baselines. As we have noted, spatial structure affects network structure by adding heterogeneity to
edge probabilities, and by creating correlations among those probabilities (an effect which is distinct from conditional dependence
among edges, as in Pattison and Robins (2002)). Nevertheless, the
possibility exists that such changes will have only a limited impact
on the global properties of the resulting networks. It has long
been known that many global network properties are sharply constrained by basic factors such as size and density (Mayhew and
Levinger, 1976; Anderson et al., 1999; Faust, 2007) and analyses such as those of Watts and Strogatz (1998) show that highly
structured networks can behave much like random graphs in
certain respects. Apart from their substantive adequacy, simple
random graph models are also useful as baselines (Mayhew, 1984a)
against which to compare the behavior of more complex models. To compare the behavior of networks generated under the
spatial Bernoulli model with their homogeneous counterparts, we
therefore constructed a paired comparison sample to our 1600
spatially conditioned networks. For each of our spatially conditioned networks, we drew a single conditional uniform graph (CUG)
with identical size and density to that in the original sample. This
resulted in a sample of equal size to the original, whose corresponding members had the same size and densities (and, by extension,
mean degrees) as the original networks, but which were free of
spatial structure. To assess the ways in which space distorts global
structure in the present context (above and beyond density), we
compare the distributions of several graph-level indices (GLIs) on
the spatial and CUG samples.
Fig. 6 summarizes the relationships between global properties of the spatial and uniform networks, as captured by several
standard graph-level indices. Each panel shows case pairs, with
vertical and horizontal coordinates respectively indicating GLI values for the spatially conditioned and CUG networks. While all of
the selected GLIs show clear differences between the models, the
nature and extent of the deviations vary. The top left panel of Fig.
6, for instance, shows the standard deviation of the degree distribution for each simulated pair. Although the mean degrees in each
case are constrained to be equal, their variations are not: as can
clearly be seen, virtually all of the spatial networks are well above
the 45-degree line, indicating that the degree distributions under
the spatial models are substantially more variable (and, in practice,
more right-skewed) than their random counterparts. Interestingly,
this amplication of variability appears to be quite systematic,
with degree standard deviation in the spatial model scaling as
approximately the 5/3 power of the random baseline (R2 = 0.96).
Moreover, we see that this relationship appears to be generally
homogeneous with respect both to the choice of SIF and to the
89
micropopulation model. This would seem to imply that the form of

the variance increase stems from a relatively robust property of the
spatial model, rather than from any region-specic geographical
features.
While degree variance is substantially and systematically different in the spatial and uniform cases, mean core number provides
an example of a GLI that behaves fairly similarly in both cases.
As the top-right panel shows, mean core number in the spatial
sample is closely and linearly related to mean core number in
the uniform sample (R2 > 0.99), with the observations deviating
only slightly from the 45-degree line. Although there is clearly
a systematic difference the spatial models have slightly higher
mean core numbers when the baseline is low, and slightly lower
mean core numbers when the baseline is high7 this difference
is small compared to the overall variation in the statistic. Intuitively, we may understand strong correlation as arising from the
fact that mean core number tends to be robustly related to degree,
but that spatial correlation makes it slightly easier to assemble
low-order cores (by grouping edges together in space) while also
making it slightly more difcult to assemble high-order cores (by
reducing connectivity between distant groups). As with variation in
degree, this relationship appears quite fundamental, and shows little evidence of depending upon the particularities of the location at
hand.
A strong contrast with the two previous cases is provided by
transitivity, i.e., the fraction of completed two-paths. As the bottom left panel of Fig. 6 shows, there is no clear relationship between
transitivity in the CUG baselines and transitivity in the spatial networks, and the correlation between the two, while signicant, is
extremely small ( = 0.085, p = 0.0006). Indeed, this is one of the
most substantial areas of divergence between the models, with
transitivity being almost entirely absent in the baseline networks
(median = 0, IQR = 3.0e5) and very substantial under the spatial
model (median = 0.28, IQR = 0.32). That the spatial model produces
a higher rate of triadic clustering is easily understood (indeed, this
was a major motivation for the latent space model of Hoff et al.,
2002), but the quantitative extent of the difference is nevertheless
dramatic. Here, we thus have an example of a network property
for which there is not only great divergence from the baseline but
also for which the baseline model appears to have little predictive
value.
Finally, we consider degree centralization, a case which lies
between the previous extremes. The measure we employ here is the
size and density normalized degree centralization of Butts (2006b),
which reduces the impact of scale and mean degree factors on the
centralization score. Since high levels of degree centralization in
large networks require the assembly of extremely large stars, it
is not surprising that levels are low overall; nevertheless, we see
that centralization is on average slightly higher in the spatial versus
non-spatial networks, a tendency which reduces as the baseline
centralization grows. The relationship here is clearly weaker than
in the rst two cases (R2 of 0.96 in loglog scale, but only 0.88 on
the original scale), suggesting a greater role for idiosyncratic effects.
On the other hand, the overall relationship is still strong enough to
indicate a close connection between the baseline behavior and the
behavior of the spatial model.
To summarize, we nd overall that there are fairly systematic
relationships between the overall behavior of the spatial model
and the behavior of the corresponding baselines implying that
one of the more important properties of the spatial model in any
given context is its expected density, a point to which we will later
return. This does not imply that the features of networks drawn
7
Note that this is not a regression artifact; as the high R2 suggests, the result does
not reverse when one regresses the CUG score on the spatial model score.
90
Fig. 6. Graph-Level properties, spatial models versus paired CUG baselines. Each point represents a single simulation outcome (location by SIF by microdistribution), with
color indicating choice of microdistribution. 45-Degree line indicates equality; red lines, where applicable, show least squares prediction of spatial from baseline GLI values.
Inset in last panel shows variables in loglog scale. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of the article.)
from the spatial model are necessarily similar to those of random graphs with equivalent dyadic characteristics (although they
are somewhat in the cases of mean core number and degree centralization), but that the spatial model deviates from the baseline
model in a quantitatively predictable way. That such relationships
would arise across locations of such highly variable structure is
an interesting nding, and suggests an avenue for gaining further
theoretical leverage. On the other hand, it is also true that some
network properties differ greatly between the spatial and uniform
cases, with little linkage between the two. Transitivity seems to be
a clear example of this, having behavior in the spatial context that
is essentially decoupled from the size/density controlled baseline.
As we shall see, however, this does not mean that transitivity is
wholly unpredictable. Rather, the way in which it arises in spatial

networks is rather different from its behavior in uniformly mixed
graphs.
3.2.1.2. Aggregate geography and GLIs. As the baseline comparison

above suggests, many aspects of our simulated networks seem to
be well-ordered, and their global behavior surprisingly predictable
from baseline properties. Given this, we may be inclined to ask
whether we can predict some aspects of global network structure
from aggregate properties of the regional geography on which they
are based (without resorting to detailed knowledge of the population surface). As it happens, there are a number of GLIs which can
be at least roughly predicted from basic locational characteristics,

several of which we summarize here.
Across scenarios, mean degree is strongly related to the mean
(spatial) nearest-neighbor distance. Indeed, a simple power law
model based on this alone (with an interaction for choice of SIF)
explains approximately 79% of the variance in mean degree. This
may be understood by a sparse/local approximation, in which it
is assumed that the probability of being tied to higher-order neighbors is small compared to the probability of being tied to the nearest
neighbor; in the limit, mean degree under this model becomes
determined by the nearest-neighbor distance (and takes an approximate power law form because of the choice of SIF). An alternative
approach is to assume that degree will scale with the local population density, and that this will in turn depend upon the degree of
global clustering (in the sense that the larger the regions of empty
space, the higher the population density must be in the remaining areas). One natural measure of the latter is the median of the
F (or empty space) function (Ripley, 1988), which is the median
radius that one must go from a randomly selected coordinate within
the sample area until one encounters a vertex. A power law model
incorporating population density, the median empty space function, and an interaction for SIF choice explains 96% of the mean
degree variance (93% on log scale), while still being quite parsimonious. This model is shown in Table 2. As we saw earlier that
mean core number generally grows linearly with degree (a theoretical result which we will revisit below), we may expect this
model to predict the former fairly well; in fact, using identical predictors on log mean core number (not shown) yields an R2 of 0.93.
Thus, both local cohesion and degree are fairly well explained by a
combination of macro-clustering and population density.
Given the ability to explain mean degree from basic geographical characteristics, density is explained even more directly. First,
we note from the earlier mentioned identity that density must be
equal to d/(N 1); since population varies here by three orders of
magnitude, while mean degree varies by less than a single order of
magnitude, the simple relation 1/N is expected to be fairly predictive. Indeed, this is the case: the correlation between density and
the inverse of population is approximately 0.96 (R2 = 0.92), and we
can thus immediately account for most of the variation in density by
population alone. This prediction can be improved upon by using
what we have already discovered regarding mean degree. Combining the predictors of Table 2 with the logged population count
yields an adjusted R2 of over 0.99 for logged density, accounting for
over 99% of the variance in the statistic in both the logarithmic and
original scale. (Because of the transparent similarity of this model
to the degree model, we omit it here.)
We saw earlier that transitivity was a graph property whose
behavior in the spatial case was very far from and generally
unrelated to the conditional uniform baseline. This does not
mean, however, that transitivity is unpredictable. Intuitively, we
should expect transitivity to be increasing in overall population
density, and decreasing with nearest neighbor distance. A simple
log-transformed model with these terms (and an interaction for
SIF) explains just over 89% of the variance in (log) transitivity, but
the effects are not exactly as one would expect: while population
density does correlate positively with transitivity, nearest neighbor distance is positive as well! The solution to this apparent puzzle
appears to lie in the fact that graphs with short mean nearest neighbor distances provide more opportunities for creating bridging ties,
thereby generating intransitive two-paths. (This intuition is reinforced by the observation that transitivity is lower under the Social
Friendship SIF, precisely because this more spatially diffuse model
allows ego to connect to relatively distant alters that are thereby
unlikely to be tied to one another.) Allowing an interaction term
between density and nearest neighbor distance boosts the R2 to
93% (92% on original scale) without adding an excessive number
91
of parameters (though interpretation of individual parameters is

difcult due to the multiplicative effects). We provide this more
complex (but conceptually similar) model in Table 3.
When networks are embedded in space, we may note that they
acquire additional properties that are due neither to network structure nor space alone, but rather to their superposition (Butts, 2002).
One simple example of such a property is the mean edge length,
meaning the average spatial distance between endpoints of a randomly selected edge. Intuitively, we may expect mean edge length
to scale with the characteristic length of each location (i.e., the
square root of the location area), and a simple log-scale model with
this term and an SIF interaction accounts for 86% of the variance in
the log edge length statistic. Interestingly, population provides an
even better model, leading to an R2 of just under 0.95. (See Table 4)
This latter effect may be due to the fact that population tends to ll
space unequally, and that one tends to get substantial numbers of
actors at long distances (and hence the opportunity for long edges)
for MSAs with high population counts.
To summarize the forgoing, many global properties of the networks formed under the spatial model appear to be predictable
from basic characteristics of the underlying geography. Choice
of SIF is also of critical importance here, unlike in the case of
spatial/baseline comparisons, as it governs the nature of the interaction between the geographic variables and the properties of the
resulting network. On the other hand, choice of population microdistribution does not appear in any of the above models, having
been found to add little or no explanatory power in any of the above
cases. This is consistent with the notion that, while low-level geographical details may affect those in particular parts of the network,
the global features of regional scale networks are driven more by
general properties of the population surface than by its intimate
details.
3.2.1.3. Internal relationships among GLIs. While there is a direct
relationship between geographical variables and GLIs, it is also
important to note that GLIs are also related to each other this can
in some cases prove to be a useful predictive device in its own right
(i.e., constraining one graph feature allows one to predict another),
and it can also provide insight into the factors that drive the creation of network structure. Before proceeding to a consideration of
within-location heterogeneity, then, we summarize some particularly noteworthy relations among global properties for networks in
our simulation sample.
We begin by considering mean degree. As with other (nonspatial) networks, we nd that mean degree is itself a powerful
predictor of other structural properties. Notably, the mean core
number correlates with mean degree at greater than 0.99, with the
relationship being strongly linear. (See Fig. 7, top left panel.) Likewise, the marginal probability of belonging to the 2-core (and thus
to a bicomponent) is essentially determined by mean degree (the
logit of the 2-core membership probability correlates with the log
mean degree at greater than 0.99). These effects are qualitatively
similar to what one observes in homogeneous random graph models, despite the dramatically inhomogeneous structure induced by
geography. We interpret this as being consistent with the notion
that, at large levels of aggregation, much of the impact of geography
is on its determination of baseline network parameters; although
spatial structure clearly induces many other non-random features (e.g., triadic closure), the baseline effects are often powerful
enough to constrain many aspects of network structure. (See also
similar arguments by Mayhew (1984a), Butts (2001), Faust (2007),
and related results shown above.)
We saw earlier that conditioning on mean degree, the standard deviation of degree increased at an accelerating rate with
the expected baseline standard deviation. The top right panel of
Fig. 7 shows that the standard deviation of degree likewise grows
92
Table 2
Regression of logged mean degree on population density (PopDen) and median empty space function (FFunMed), with an interaction for model type.
(Intercept)
FriendSIF
log(PopDen)
log(FFunMed)
FriendSIF log(PopDen)
FriendSIF log(FFunMed)
Estimate
Std. error
t-Value
Pr (> |t|)
0.47
2.37
0.11
0.13
0.28
0.25
0.035
0.050
0.004
0.006
0.005
0.008
13.39
47.26
30.78
22.45
55.74
30.56
<2e16***
<2e16***
<2e16***
<2e16***
<2e16***
<2e16***
Residual standard error: 0.1717 on 1594 degrees of freedom.

Multiple R2 : 0.9314, adjusted R2 : 0.9311.
F-statistic: 4326 on 5 and 1594 DF, p-value: <2.2e16.
***
p < 0.001.
**
p < 0.01.
*
p < 0.05.
Table 3
Regression of logged transitivity on population density (PopDen), mean nearest neighbor distance (NNDist), and SIF, with all pairwise interactions.
(Intercept)
log(NNDist)
log(PopDen)
FriendSIF
log(NNDist) log(PopDen)
log(NNDist) FriendSIF
log(PopDen) FriendSIF
log(NNDist) log(PopDen) FriendSIF
Estimate
Std. error
t-Value
Pr (> |t|)
2.75
0.84
0.09
5.11
0.05
1.35
0.18
0.07
0.141
0.060
0.013
0.199
0.004
0.085
0.019
0.006
19.56
14.00
6.99
25.71
12.53
15.89
9.53
12.31
<2e16***
<2e16***
4.13e12***
<2e16***
<2e16***
<2e16***
<2e16***
<2e16***

F-statistic: 3247 on 7 and 1592 DF, p-value: < 2.2e16.
***
p < 0.001.
**
p < 0.01.
*
p < 0.05.
systematically (but sublinearly) with the mean degree itself

(R2 = 0.96). While the degree standard deviation grows more slowly
in these models than the mean, it nevertheless grows more rapidly
than the square root of the mean (dotted line), and hence is inconsistent with a Poisson distribution. We will return to this issue in
the next section, when we directly evaluate the degree distributions
for the full set of networks.
We saw previously that transitivity was related to many of the
same geographical properties that predict mean degree, and the
relationship between the two GLIs is claried in the lower left panel
of Fig. 7. Transitivity falls smoothly with mean degree (R2 = 0.86),
an effect related (as noted earlier) to the fact that the same circumstances giving rise to high degree also give rise to an increase
in local bridges, and thus to intransitivity. The transitivity/mean
degree relationship helps to clarify the fact that, for these networks,
global transitivity is a feature of highly fragmented, extremely local
structures in which only small, spatially proximate clusters tend to
be well-connected.
Finally, the upper right panel of Fig. 7 shows that a moderately strong relationship exists between normalized degree
centralization and density (R2 = 0.98 on log scale). It is noteworthy in this regard that a similar relationship does not hold for mean
degree (R2 = 0.13 on log scale), bearing in mind that density is more
heavily driven by population size than by mean degree in this sample. It should also be borne in mind that the centralization measure
employed here is already normalized for density as well as graph
size, and hence the observed density relationship is more subtle
than it might appear. We note in this regard that the highest density graphs are simultaneously those that are small and that have
higher mean degree. These graphs will be more likely to, on the
one hand, produce high-degree outliers (since they have relatively
higher degree variance), and, on the other, to have maximum raw
centralization scores (with respect to which the measure is normalized) that are small enough that the normalized measure is not
damped heavily towards zero. Complex measures such as centralization thus require careful interpretation (even when adjusted
for baseline effects), no less so for these than for other network
models.
In summary, strong and systematic relationships exist between
GLIs within this sample. Some appear to be driven by baseline
Table 4
Regression of logged edge length on logged population count (Pop) and SIF, with a population/SIF interaction.
(Intercept)
log(Pop)
FriendSIF
log(Pop) FriendSIF
Estimate
Std. error
t-Value
Pr (> |t|)
0.57
0.10
0.83
0.18
0.051
0.005
0.072
0.007
10.97
20.72
11.60
26.98
<2e16***
<2e16***
<2e16***
<2e16***

F-statistic: 9461 on 3 and 1596 DF, p-value: < 2.2e16.
***
p < 0.001.
**
p < 0.01.
*
p < 0.05.
93
Fig. 7. Relationships among various graph-level properties, simulated spatial networks. Each point represents a single simulation outcome (location by SIF by microdistribution), with color indicating choice of microdistribution. Red lines, where applicable, show least squares prediction of vertical axis GLI from horizontal axis GLI. Insets show
variables in log-log scale. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of the article.)
effects; their presence here reinforces the observation that even

strong underlying heterogeneity does not immediately overcome
combinatorial considerations. Other relationships, however (like
that between mean degree and the degree variance) are clearly
the result of spatial effects per se. It is interesting to note that
neither SIF choice nor microdistribution model were found to be
important mediators of the relationships among GLIs for these networks, even for those with an obviously spatial character; this
suggests at the very least that the relationships are reasonably
robust to detailed modeling assumptions, and perhaps that they
are reective of very general effects arising in a range of settings
(spatial and otherwise). While these simple often linear or multiplicative relationships among global network properties thus
demonstrate an order arising from a heterogeneous foundation,
they do not tell the whole story. As we shall now show, the detailed
structure of these networks clearly shows the impact of geographical factors.

3.2.2. Within-location heterogeneity
Having shown consistent behavior in the cross-location context, we now look within each network (using the smaller sample
of 64 example structures) to examine the inuence of geographical heterogeneity on local network structure. Consistent
with the arguments of Section 2, inspection of the 64 networks
reveals that local properties such as degree and core number
vary systematically across space. As Figs. 8 and 9 illustrate, the
frequency of persons with high degree and high-order k-core
members increases as one approaches regions of high population, with those in relatively sparsely populated areas tending to
have both lower degree and lower levels of local cohesion. These
94
Fig. 8. Spatial variation in predicted Degree (by SIF and layout model), Navajo County, AZ. Vertex color indicates degree, with bluer colors indicating higher greater numbers
of ties. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of the article.)
relationships are as we would expect from rst principles. While

there is a strong spatial relationship with both degree and core
number (both tracking the population surface), it should be noted
that the highly clustered nature of population at the micro-level
(in part associated with family structure) helps to ensure that
some persons of at least moderate degree can be found even in
outlying areas (and some of low degree in areas of high population). In practice, this mixing would mean that spatial degree
variation could easily go unperceived by residents of a given
area, even where the mean differences between dense and sparse
regions are fairly substantial. Core number is even less likely to
be perceived, since this would require extensive knowledge of the
local network (and the ability to carry out the necessary calculations). While these effects may be invisible to residents, they
could nonetheless be veried by appropriate empirical studies
(though the authors know of no extant data set that speaks to this
issue).
Also in line with the arguments of Section 2, the wildly unequal

distribution of population across space leads to dramatic differences in local connectivity and tie volume. This is graphically
illustrated in Fig. 10, which shows ties among individuals in
blocks near the center of Cookeville, TN (quasi-random distribution,
Friendship SIF). While activity is present throughout the region, the
intense clustering of persons in blocks like that near the center of
the gure creates a corresponding social cluster whose members
have both higher mean degree and who are on average more cohesively connected than those in nearby blocks. Even at scales on the
order of 1 km, we thus expect to see substantial heterogeneity in
structural characteristics that are driven in part by geographical
variation.
As the above example suggests, subgraphs in a spatial context
have a dual existence: they can be considered on the one hand in
terms of their network properties, and on the other in terms of
the spatial positions of their members. In this light, we note that
Table 5
AICC selected models for degree distribution, by location, SIF, and placement model.
Bristol Bay, AK
Golden Valley, MT
Esmeralda, NV
Yakutat, AK
Choctaw, MS
Cheyenne, NE
Quay, NM
White Pine, NV
Lawrence, KS
Cookeville, TN
Idaho Falls, ID
Navajo, AZ
Honolulu, HI
Hartford, CT
Rochester, NY
Salt Lake City, UT
Freeman/uniform
Festinger/uniform
Freeman/quasi
Festinger/quasi
NB
NB
NB
NB
NB
NB
NB
NB
NB
W
NB
NB
NB
W
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
W
NB
NB
NB
W
NB
NB
NB
P
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
G, Geometric; NB, Negative binomial; P, Poisson; W, Waring; Y, Yule.
95
Fig. 9. Spatial variation in predicted core number (by SIF and layout model), Cookeville, TN MSA. Vertex color indicates maximum core membership, with bluer colors
indicating membership in higher-order cores. (For interpretation of the references to color in this gure legend, the reader is referred to the web version of the article.)
the convex hull of the set of vertices in a particular subgraph provides an intuitive notion of the region covered by that subgraph,
as illustrated in Fig. 11. The gure graphically illustrates the emergence of large cohesive subgroups (here, members of high-order
cores who are themselves biconnected) within the Cookeville, TN
case. As argued in Section 2, such groups should develop relatively suddenly when a sufciently large area exceeds the requisite
threshold density; the location of large cores covering the highdensity regions of the map is consistent with this behavior. Such
spatially large cohesive sets are of potential interest for theories
such as those of Sampson et al. (1997), which relate to the ability of
social groups to monitor and control activities within a given area.
Models of the kind studied here suggest a relatively sharp boundary between the conditions under which such cohesion is feasible,
and those under which it is not. Such boundaries may account
in part for the frequently voiced sense of qualitative difference
between social interactions in cities and those in sparsely populated
environments.
While the detailed structure of the simulated networks reveals
substantial variation in both positional and group characteristics,
there are nevertheless strong similarities across cases. Fig. 12 shows
the marginal degree distributions for all 64 networks, each representing the aggregate effect of the interaction of population surface
and SIF across vertices. While the proles of Fig. 12 obviously differ
96
Fig. 10. Detail of edge structure (quasi-random placement, Friendship SIF) for a portion of the Cookeville, TN MSA. Vertices are shaded by core number, from red (minimum)
to violet (maximum); dark lines indicate census block boundaries. (For interpretation of the references to color in this gure legend, the reader is referred to the web version
of the article.)
in some respects, there are also some similarities (not the least
of which being a relatively long upper tail). Are these similarities only a matter of appearance, or do they indicate a common
underlying functional form? To assess this, we attempted to t
geometric, negative binomial, Poisson, Waring, and Yule distributions to each of the depicted degree distributions (models t via
degreenet (Handcock, 2003) using maximum likelihood). After
tting all ve distributions to each network, we select the model

that best approximates the observed distribution (in the sense of
minimizing the expected KullbachLeibler distance) using the AICC
statistic. The resulting classication of network by model is provided in Table 5. The outcome of this analysis is overwhelmingly
clear: in 59 out of 64 cases the preferred model is the negative binomial, with the long-tailed Waring distribution preferred in four and
Fig. 11. Spatial structure of cohesive components, Cookeville TN MSA (uniform placement, Friendship SIF). Shaded regions indicate convex hulls of membership locations
for biconnected sets of k-core members, with pink shading indicating 2-cores, and green indicating 3-cores. Right-hand panel shows detail of dotted area. (For interpretation
of the references to color in this gure legend, the reader is referred to the web version of the article.)
97
Fig. 12. Marginal degree distributions by location, SIF, and placement model. Friendship model distributions are shown in blue, interaction model distributions in black;
solid lines indicate uniform placement, with quasi-random placement in dotted lines. (For interpretation of the references to color in this gure legend, the reader is referred
to the web version of the article.)
the Poisson preferred in only 1. The Waring cases seem to be associated with the Hartford, CT and Cookeville, TN MSAs under the
Friendship SIF, and may reect particularly high levels of heterogeneity in these locations. These possible exceptions aside, the vast
majority of cases can be seen to be well-approximated by distributions of the same form, despite differing in population size and land
area by several orders of magnitude.
Turning to core number, we note in the marginal distributions
of Fig. 13 the same combination of family resemblance and difference in detail seen earlier in Fig. 12. As before, we attempt to assess
the presence of a common underlying distributional form by tting

models to each distribution, selecting that chosen by the AICC. The
results of this process are shown in Table 6. Once again, we nd that
the negative binomial is overwhelmingly preferred, with the geometric distribution (itself a special case of the negative binomial)
favored in two cases (both Hartford, CT Interaction SIF models) and
the Waring favored in one. For core number, as for degree, then, the
unique pattern of variation in each individual population surface
nevertheless combines to generate a consistent family of marginal
distributions.
Table 6
AICC selected models for core number distribution, by location, SIF, and placement model.
Bristol Bay, AK
Golden Valley, MT
Esmeralda, NV
Yakutat, AK
Choctaw, MS
Cheyenne, NE
Quay, NM
White Pine, NV
Lawrence, KS
Cookeville, TN
Idaho Falls, ID
Navajo, AZ
Honolulu, HI
Hartford, CT
Rochester, NY
Salt Lake City, UT
Freeman/uniform
Festinger/uniform
Freeman/quasi
Festinger/quasi
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
G
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
W
NB
NB
NB
G
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
NB
G, Geometric; NB, Negative Binomial; P, Poisson; W, Waring; Y, Yule.
98
Fig. 13. Marginal core number distributions by location, SIF, and placement model. Friendship model distributions are shown in blue, interaction model distributions in
black; solid lines indicate uniform placement, with quasi-random placement in dotted lines. (For interpretation of the references to color in this gure legend, the reader is
referred to the web version of the article.)
To summarize our ndings regarding within-region/withinnetwork variation, the results from our simulated networks closely
follow our a priori expectations. Degree and core number vary
with the population surface, and cohesively connected subgroups
appear over regions with systematically above-threshold density.
While all of this suggests that many local structural properties
will vary greatly both within and across regions, this variation is
also systematically structured. In addition to the above patterns,
we nd that both degree and core number for these networks
are well-modeled by negative binomial distributions (though the
parameters of those distributions clearly vary by location and SIF).
Although it is not clear how robust this result is to the imposition
of other (currently unmodeled) factors, its prevalence here leads to
the speculation that the negative binomial degree and core number
distributions may serve as easily falsiable signatures for a spatial
Bernoulli process. It is interesting in that light to note that the negative binomial has been found to provide a reasonable t to at least
some empirical data sets (see, e.g. Hamilton et al., 2008), suggesting
that this pattern is not beyond the bounds of plausibility.
4. Discussion and conclusion

Assuming a simple, maximum entropy graph distribution constrained by empirically obtained marginal distance/tie probability
relationships for two relations, we nd that spatial variability
should indeed exert substantial inuence on network structure
at the settlement level. The highly uneven density of population
within our sample areas results in lumpy networks that are
characterized by regions of differential local connectivity, spatially
correlated gradients of expected degree and core number, and other
such properties. At small spatial scales, then, we predict that the

character of the local structural environment will for these types
of relations depend heavily on local population distribution.
While spatial heterogeneity does induce substantial withinnetwork heterogeneity, we also observe that geography drives
many aggregate network properties in a predictable way. For relations like those modeled here, aggregate mean degree, edge length,
and local clustering can be well-predicted by properties such as the
mean nearest-neighbor distance, together with SIF-specic factors.
The nearest-neighbor distance is itself driven in large part by land
area and total population size, though microdistributional factors
do play a role. Taken together, this implies that, for these sorts of
relations, it should be possible to predict differences in a number
of aggregate structural properties from fairly basic features of the
underlying social geography.
Looking forward, we note that while the present investigation
leaves out many factors e.g., endogenous clustering, differential
mixing due to age or race, and degree constraints, to name a few
it nevertheless includes a far more detailed consideration of spatial
factors than has been pursued in past studies. In that light, it is interesting to observe that many features taken as generically indicative
of complex endogenous processes (including long-tailed degree
distributions, differential mixing, and triadic clustering) arise here
in natural way without recourse to more complex assumptions.
While the existence of such processes is well-documented and
uncontroversial, their relative importance in determining the properties of large-scale networks is less clearly established. To that
end, it would be productive to undertake more systematic empirical research to identify the limits of simple properties such as
spatial marginals for structural prediction. The spatial Bernoulli
graphs are obviously an approximation to a more complex reality,
but may still be adequate for many purposes. If so, their simplicity
and tractability (theoretical, computational, and inferential) have
much to recommend them.
In like vein, we would suggest that in settings for which spatial Bernoulli graphs are inadequate, augmenting the base model
(per Eq. (2)) with covariate effects preserving Bernoulli structure
(e.g., age or race mixing) rather than sources of dependence among
edges should be considered as an initial strategy (at least for modeling of large-scale networks). Although more complex than purely
spatial models, such hybrids retain many of the scalability and theoretical tractability advantages exploited here. One natural avenue
for such expansion is via the use of Blau-space models (McPherson,
1983, 2004) that employ a notion of distance over a combined sociophysical space. (Some initial steps in this direction have been taken
by Hipp and Perrin (2009).) One major obstacle to current progress
in this area is a lack of high-quality network data sets containing
sufcient information on both geography and demographic characteristics to permit estimation of a joint socio-physical SIF. As such
data becomes available for multiple relational types, it will be possible to substantially expand the range of questions that can be
asked within the spatial network paradigm.
Another obstacle to further theoretical progress is a lack of
detailed population data on the co-evolution of social ties and residential mobility. In this paper, we have focused on the problem
of instantaneous prediction: given the (current) distance structure, predict the contemporaneous properties of an associated
social network. In so doing, we neither make nor require assumptions regarding the dynamic processes that produce this joint
socio-geographic structure so long as we know how space is
instantaneously related to social relationships, we can predict the
latter from the former. Many interesting questions, however, relate
to these underlying processes, and a detailed understanding of
them would advance current knowledge in many respects. For
instance, we here take the SIF as given (estimated from prior data),
but the marginal relationship it describes clearly arises from a
combination of (possibly tie-inuenced) movement of persons in
space, and (possibly spatially inuenced) formation and dissolution of social ties. A richer understanding of these mechanisms
could potentially allow for the prediction of spatial interaction
functions themselves, as well as prediction of the circumstances
in which such functions might vary or change. Likewise, sudden
perturbations to normal mobility patterns (e.g., displacement following wars or natural disasters) may produce short-term changes
in socio-geographic structure that are poorly predicted by comparative statics (i.e., equilibrium structure before and signicantly
after the event). Modeling such shocks requires knowledge of how
rapidly ties decay following relocation, the rate at which new ties
form in response to this relocation (and to whom), and selective
inuences of tie acquisition and loss on secondary mobility. These
are complex phenomena, which in our opinion require a much
deeper understanding of social dynamics than is currently available. However, the rst step in developing this understanding will
clearly be the design of studies that are sensitive to both spatial and
temporal concerns.
Finally, we close by reiterating some simple predictions that,
seeming to be robustly present in the cases studied here, lend
themselves to empirical evaluation. First, we predict a positive correlation between local population density and both mean degree
and core number over scales that are at least comparable to the
relevant SIF. Second, we predict transitivity far in excess of CUG
baselines, declining globally in mean degree, increasing in nearestneighbor distance, and declining in SIF tail weight. Third, we predict
sudden changes in the formation of large, cohesively connected
subgroups with population density, with such groups transitioning
from small and rare to large and common as a threshold function of local population. Finally, we predict that internal density
99
for areas of equivalent geometry and population distribution will

be approximately constant in total population (with the solution to
the degree bound problem of Mayhew and Levinger (1976) tending
to be resolved through changes in population distribution, including articial elevation, rather than through density change). None
of these predictions seems to us self-evidently true, and all could
be far from the mark. Systematic evaluation of each, however, for
a range of networks, could tell us much about how geographical
variability relates to network structure.
References
Anderson, B.S., Butts, C.T., Carley, K.M., 1999. The interaction of size and density with
graph-level indices. Social Networks 21 (3), 239267.
Bivand, R.S., Pebesma, E.J., Gmez-Rubio, V., 2008. Applied Spatial Data Analysis with
R. Springer, New York.
Bossard, J.H.S., 1932. Residential propinquity as a factor in marriage selection. American Journal of Sociology 38, 219244.
Brakman, S., Garretsen, H., Van Marrewijk, C., Van Den Berg, M., 1999. The return of
Zipf: towards a further understanding of the rank-size distribution. Journal of
Regional Science 39 (1), 183213.
Burian, S.J., Brown, M.J., Velugubantla, S.P.,2002. Building height characteristics in
three U.S. cities. In: Fourth Symposium on Urban Environment. American Meteorological Society, Norfolk, VA, pp. 129130.
Butts, C.T., 2001. The complexity of social networks: theoretical and empirical ndings. Social Networks 23 (1), 3171.
Butts, C.T., 2002. Spatial Models of Large-scale Interpersonal Networks. Carnegie
Mellon University, Doctoral Dissertation.
Butts, C.T., 2003. Predictability of large-scale spatially embedded networks. In:
Breiger, R., Carley, K.M., Pattison, P. (Eds.), Dynamic Social Network Modeling and Analysis: Workshop Summary and Papers. National Academies Press,
Washington, DC.
Butts, C.T., 2006a. Curved exponential family parameterizations for spatial network
models. In: Presentation to the 26th Sunbelt Network Conference (INSNA).
Butts, C.T., 2006b. Exact bounds for degree centralization. Social Networks 28 (4),
283296.
Butts, C.T., 2008. Social network analysis with SNA. Journal of Statistical Software
24 (6).
Butts, C.T., 2010. Bernoulli graph bounds for general random graphs. Technical
Report MBS 10-07, Irvine, CA.
Butts, C.T., Acton, R.M., 2011. Spatial modeling of social networks. In: Nyerges, T.,
Couclelis, H., McMaster, R. (Eds.), The SAGE Handbook of GIS and Society. SAGE
Publications (Chapter 12).
Carley, K.M., 2002. Computational organizational science and organizational engineering. Simulation Modeling Practice and Theory 10, 253269.
Daraganova, G., Pattison, P., 2007. Social networks and space. In: Presentation to the
2007 International Workshop on Social Space and Geographical Space.
Faust, K., 2007. Very local structure in social networks. Sociological Methodology 37
(1), 209256.
Festinger, L., Schachter, S., Back, K., 1950. Social Pressures in Informal Groups. Stanford University Press, Stanford, CA.
Freeman, L.C., Freeman, S.C., Michaelson, A.G., 1988. On human social intelligence.
Journal of Social and Biological Structure 11, 415425.
Gentle, J.E., 1998. Random Number Generation and Monte Carlo Methods. Springer,
New York.
Hgerstrand, T., 1967. Innovation Diffusion as a Spatial Process. University of Chicago
Press, Chicago.
Hamilton, D., Handcock, M.S., Morris, M., 2008. Degree distributions in sexual networks: a framework for evaluating evidence. Sexually Transmitted Diseases 35
(1), 3040.
Handcock, M.S., 2003. degreenet: Models for Skewed Count Distributions Relevant
to Networks. Seattle, WA. Version 1.0.
Handcock, M.S., Hunter, D.R., Butts, C.T., Goodreau, S.M., Morris, M., 2008. statnet:
software tools for the representation, visualization, analysis and simulation of
network data. Journal of Statistical Software 24 (1).
Handcock, M.S., Raftery, A.E., Tantrum, J.M., 2007. Model based clustering for social
networks. Journal of the Royal Statistical Society, Series A 170, 301354.
Haynes, K.E., Fotheringham, A.S., 1984. Gravity and Spatial Interaction Models. Sage,
Beverly Hills, CA.
Hipp, J.R., Perrin, A.J., 2009. The simultaneous effect of social distance and physical distance on the formation of neighborhood ties. City and Community 8 (1),
525.
Hoff, P.D., Raftery, A.E., Handcock, M.S., 2002. Latent space approaches to social
network analysis. Journal of the American Statistical Association 97 (460),
10901098.
Latan, B., Liu, J.H., Nowak, A., Bonevento, M., Zheng, L., 1995. Distance matters:
physical space and social impact. Personality and Social Psychology Bulletin 21
(8), 795805.
Mayhew, B.H., 1984a. Baseline models of sociological phenomena. Journal of Mathematical Sociology 9, 259281.
Mayhew, B.H., 1984b. Chance and necessity in sociological theory. Journal of Mathematical Sociology 9, 305339.
100
Mayhew, B.H., Levinger, R.L., 1976. Size and density of interaction in human aggregates. American Journal of Sociology 82, 86110.
McPherson, J.M., 1983. An ecology of afliation. American Sociological Review 48,
519532.
McPherson, J.M., 2004. A blau space primer: prolegomenon to an ecology of afliations. Industrial and Corporate Change 13, 263280.
McPherson, J.M., Smith-Lovin, L., Cook, J.M., 2001. Birds of a feather: homophily in
social networks. Annual Review of Sociology 27, 415444.
Pattison, P.E., Robins, G.L., 2002. Neighborhood-based models for social networks.
Sociological Methodology 32, 301337.
Ripley, B.D., 1988. Statistical Inference for Spatial Processes. Cambridge University
Press, Cambridge.
Sampson, R.J., Raudenbush, S.W., Earls, F., 1997. Neighborhoods and violent crime:
a multilevel study of collective efcacy. Science 277, 918923.
Snijders, T.A.B., 2002. Markov chain Monte Carlo estimation of exponential random
graph models. Journal of Social Structure 3 (2).
Wasserman, S., Faust, K., 1994. Social Network Analysis: Methods and Applications.
Cambridge University Press, Cambridge.
Watts, D.J., Strogatz, S.H., 1998. Collective dynamics of small-world networks.
Nature 393, 440442.
West, D.B., 1996. Introduction to Graph Theory. Prentice Hall, Upper Saddle River,
NJ.
White, D.R., Tambayong, L., Kejzar, N., 2008. Oscillatory dynamics of city-size distributions in world historical systems. In: Modelski, G., Devezas, T., Thompson,
W.R. (Eds.), Globalization as an Evolutionary Process: Modeling Global Change.
Routledge, London.
Zipf, G.K., 1949. Human Behavior and the Principle of Least Effort.

1 s2.0 S0378873311000517 Main PDF

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

1 s2.0 S0378873311000517 Main PDF

Caricato da

Copyright:

Formati disponibili

Social Networks 34 (2012) 82100

Contents lists available at SciVerse ScienceDirect

Geographical variability and network structure

A basic observation regarding the distribution of humans across

C.T. Butts et al. / Social Networks 34 (2012) 82100

distance and tie probability. Using this simplied framework, we

1. Spatial models of social networks

B(Yij = yij |F(Dij , )),

where Y is the (random) graph adjacency matrix, D is a matrix of

(, Dij )yij +

where Rp and t : Y  Rp are respective vectors of parameters

C.T. Butts et al. / Social Networks 34 (2012) 82100

2. Effects of spatial heterogeneity: intuitions from rst

p(j )F(Dij , )dj pY (i , )

where i , j are the locations of the respective vertices and p( j )

pY (1 , ) = kpY (1 , ).

Marginalizing over 1 gives the expected degree of vi ,

kpY (1 , )d1 ,

The in-lling principle has a number of further consequences,

C.T. Butts et al. / Social Networks 34 (2012) 82100

Fig. 3. Emergence of local connectivity on an uneven population density surface.

Taking the strong relationship of mean core number with N at

C.T. Butts et al. / Social Networks 34 (2012) 82100

C.T. Butts et al. / Social Networks 34 (2012) 82100

Bristol Bay Borough, AK

a total of 16 study locations. These are listed in Table 1. As can

Land area (km2 )

Population density (Pop/km2 )

arises as population density grows, with new households stacking

3.1.3. Spatial interaction functions

C.T. Butts et al. / Social Networks 34 (2012) 82100

data from (Festinger et al., 1950)), and can be thought of as a

a comparison of Euclidean distances for the cases used here with

C.T. Butts et al. / Social Networks 34 (2012) 82100

micropopulation model. This would seem to imply that the form of

C.T. Butts et al. / Social Networks 34 (2012) 82100

wholly unpredictable. Rather, the way in which it arises in spatial

3.2.1.2. Aggregate geography and GLIs. As the baseline comparison

C.T. Butts et al. / Social Networks 34 (2012) 82100

be at least roughly predicted from basic locational characteristics,

of parameters (though interpretation of individual parameters is

C.T. Butts et al. / Social Networks 34 (2012) 82100

Residual standard error: 0.1717 on 1594 degrees of freedom.

Residual standard error: 0.2433 on 1592 degrees of freedom.

systematically (but sublinearly) with the mean degree itself

Residual standard error: 0.3461 on 1596 degrees of freedom.

C.T. Butts et al. / Social Networks 34 (2012) 82100

effects; their presence here reinforces the observation that even

structure of these networks clearly shows the impact of geographical factors.

C.T. Butts et al. / Social Networks 34 (2012) 82100

relationships are as we would expect from rst principles. While

Also in line with the arguments of Section 2, the wildly unequal

G, Geometric; NB, Negative binomial; P, Poisson; W, Waring; Y, Yule.

C.T. Butts et al. / Social Networks 34 (2012) 82100

C.T. Butts et al. / Social Networks 34 (2012) 82100

tting all ve distributions to each network, we select the model

C.T. Butts et al. / Social Networks 34 (2012) 82100

the presence of a common underlying distributional form by tting

G, Geometric; NB, Negative Binomial; P, Poisson; W, Waring; Y, Yule.

C.T. Butts et al. / Social Networks 34 (2012) 82100

4. Discussion and conclusion

such properties. At small spatial scales, then, we predict that the

C.T. Butts et al. / Social Networks 34 (2012) 82100

where Rp and t : Y Rp are respective vectors of parameters